I discovered models have unique handwriting and matching them reveals something about how they are related

One thing I love about Claude is that it enables me to research whatever I want. I was curious about whether models have their own handwriting habits and whether they would be similar in the same family or lineage. For a small scale interpretability research I do, I collected quite a lot of data during the last 3 months which I could reuse for this purpose. I let Claude run autonomously overnight and asked it to analyze all the code I gathered to find candidate habits we could use to identify models with a high certainty. It delivered about 40 different habits that could be measured. This allowed me to plot models based on the vectors of their handwriting and calculate their similarity to the writing of other models, which turned out to closely align with how models are (probably) related per lineage in reality. Being careful here, because this is a tiny research, but assuming most labs use distillation with teacher and student models; The smaller the distance, the more likely the handwriting similarity suggest a relationship. Most notable findings: the GPT-6 models are all very closely related, while there is much more of a spread with the Claude models. The newer Sonnet and Opus models show much more variation in their handwriting (see the Halos in the interactive graph), I expect the new Fable-5.5 model to have a similar trait. Sonnet-5.5 is somehow really far removed from all the other Claudes, but this could all change when Fable-5.5 is added. Interactive graph made by Claude and hosted here: curiosity.observer/handwriting